Architectural Scalability and Attention Stabilization for Robust Impulsive Acoustic Spectrogram Classification Using Swin Transformer V2
Year: 2026
Type: article
Abstract: Impulsive acoustic event classification based on spectrogram representations remains challenging due to the non-stationary and broadband characteristics of transient signals. While convolutional neural networks (CNNs) have been widely adopted for spectrogram-based classification, their reliance on local convolutional receptive fields may limit the modeling of long-range spectro–temporal dependencies. This study investigates the architectural scalability and attention stabilization properties of Swin Transformer V2 for robust impulsive acoustic spectrogram classification. A comprehensive experimental framework is developed using twelve time–frequency representations, including Bark, Chromagram, Cochleagram, Constant-Q Transform (CQT), Delta, Delta-Delta, Log-Mel, Mel, MFCC, Spectral Contrast, Short-Time Fourier Transform (STFT), andWavelet spectrograms. Multiple Swin Transformer configurations are evaluated and compared with widely used CNN baseline architectures, including ResNet50, ResNet101, EfficientNet-B0, and InceptionResNetV2, under a unified training protocol with five-fold stratified cross-validation. Experimental results demonstrate that Swin Transformer V2 consistently achieves higher classification accuracy and macro-F1 scores than CNN baselines across diverse spectrogram representations. In addition, the hierarchical attention mechanism provides improved stability across data partitions and enhanced robustness under spectrogram resolution variability. These findings highlight the effectiveness of transformer-based architectures for modeling complex spectro–temporal structures in impulsive acoustic signals and provide insights into the architectural behavior of Swin Transformer models for spectrogram-based acoustic classification.
Source: IEEE Access
Authors: Pafan Doungpaisan, Peerapol Khunarsa
Language: English
Topic: Music and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7147477145